Meta-Chunking: Learning Efficient Text Segmentation via Logical Perception

Retrieval-Augmented Generation (RAG), while serving as a viable complement to large language models (LLMs), often overlooks the crucial aspect of text chunking within its pipeline, which impacts the quality of knowledge-intensive tasks. This paper introduces the concept of Meta-Chunking, which refers to a granularity between sentences and paragraphs, consisting of a collection of sentences within a paragraph that have deep linguistic logical connections. To implement Meta-Chunking, we designed Perplexity (PPL) Chunking, which balances performance and speed, and precisely identifies the boundaries of text chunks by analyzing the characteristics of context perplexity distribution. Additionally, considering the inherent complexity of different texts, we propose a strategy that combines PPL Chunking with dynamic merging to achieve a balance between fine-grained and coarse-grained text chunking. Experiments conducted on eleven datasets demonstrate that Meta-Chunking can more efficiently improve the performance of singlehop and multi-hop question answering based on RAG. For instance, on the 2Wiki-MultihopQA dataset, it outperforms similarity chunking by 1.32 while only consuming 45.8% of the time. Furthermore, through the analysis of models of various scales and types, we observed that PPL Chunking exhibits notable flexibility and adaptability. This is particularly relevant in knowledge-intensive tasks like open-domain question answering (Lazaridou et al., 2022). By integrating two key components: the retriever and the generator, this technology enables more precise responses to input queries (Singh et al., 2021; Lin et al., 2023). While the feasibility of the retrieval-augmentation strategy has been widely demonstrated through practice, its effectiveness heavily relies on the relevance and accuracy of the retrieved documents (Li et al., 2022; Tan et al., 2022). The introduction of excessive redundant or incomplete information through retrieval not only fails to enhance the performance of the generation model but may also lead to a decline in answer quality (Shi et al., 2023; Yan et al., 2024).